The Fourth GMIF2025 Innovation Summit (Global Memory Innovation Forum) recently wrapped up in Shenzhen. Themed "AI Applications, Innovation Empowered, "GMIF2025 represented as a gathering of leading enterprises, technical experts and industry leaders across the memory sector to explore pathways for technological evolution and emerging opportunities driven by AI. Kevin Yoon, VP and CTO of Memory Division from Samsung Electronics, delivered a keynote titled "Architecting Al Advancement: The Future of Memory and Storage", unveiling a series of cutting-edge solutions and providing a glimpse into the challenges and innovation directions of memory technologies.
Kevin Yoon, VP and CTO of Memory Division from Samsung Electronics. Credit:GMIF
Memory is Embracing an Architecture Reconfiguration with the Advent of Agentic AI Era
According to the address by Kevin Yoon, AI is evolving from Generative AI to Agentic AI and will progress toward Physical AI. The memory industry is now officially transitioning into the Agentic AI era. Characterized by independent inference and decision-making capabilities, the Agentic AI is demanded to maintain multiple states over time built on top of generative AI and invoke external tools without noticeable latency. This shift is fueling explosive growth in global data and computing workloads.
"The inference of Agentic AI requires longer periods to achieve optimal outcomes, resulting in surging needs of memory capacity," Kevin Yoon underlined, data centers are transitioning into data-intensive computing mode in a fast manner. Therefore, memory technologies goes beyond a backend underpinning to become the determinant of AI system efficiency and scalability. Legacy storage architectures are faced by exponential bandwidth demands, power consumption nearing physical limits, and increasing latency constraints, so the memory hierarchy must be restructured to meet the needs of next-generation intelligent infrastructure.
Dual-Track Advances in Memory: Samsung Mass Produces GDDR7 and Leads the CXL Ecosystem
Regarding high-bandwidth needs of AI servers, Kevin Yoon announced its major breakthroughs. Its first-ever 24Gb GDDR7 in mass production establishes deep collaborations with leading GPU partners. Engineered with leading-edge process and optimized IC architecture, it delivers transfer rate up to 42.5Gbps, a 30% enhancement in energy efficiency compared to last-gen products, making it stand out in AI training and graphics rendering.
Compute Express Link (CXL) stands as a promising solution to address the capacity limit of DRAM. Samsung, the first to debut CXL products, has already mass-produced CXL 2.0 products benefited by its technology and ecosystem development since 2021. Kevin Yoon noted that as the industry moves into the CXL 3.0 era, Samsung is developing devices that support real-time memory sharing across multiple servers. A CMM-D solution with CXL 3.1 and PCIe Gen 6.0 compatibility is scheduled for launch next year, with plans to add full-feature support such as near-memory processing to further improve flexibility and scalability in data centers.
All-Around Innovation: Advancing Performance, Density and Thermal Management
In memory sector, Kevin Yoon introduced Samsung's technology layout and product roadmap in terms of three major indicators: high performance, high density and thermal control. On the performance front, Samsung is tackling the challenges posed by interface upgrades to PCIe through coordinated design across NAND and controllers and is planning to launch its PCIe Gen6 SSD PM1763 in early 2026. This brand-new versions is expected to offer doubled performance under the power consumption limit of 25W and a 1.6-fold energy efficiency improvement, making it ideally suited to GPU-intensive AI computing scenarios.
With respect to high-density storage, Samsung is clearly ahead of the pack. It launched a 128TB U.2 SSD in 2025 and plans to roll out 1T-thickness EDSFF products between 2026 and 2027, pushing capacities to 256TB on Gen5 and 512TB on Gen6. These leaps are powered by Samsung's advanced 32-layer stacked packaging and the slim form factor of EDSFF, enabling higher density and better energy efficiency in limited space.
To address the thermal challenges of high-performance storage, Kevin Yoon explained that Samsung is pivoting from traditional air cooling to direct liquid cooling technologies. By reducing the thickness of its E1.S 8TB SSDs from 15T to 9.5T and minimizing thermal resistance between cold plates and enclosures, this will ensure system stability under full-load conditions for next-generation storage products.
New Product Category: Memory Class Storage Ushers in a Low-Latency Era
To support the rapid-access needs of AI inference on small data blocks, Samsung introduced a new category called Memory Class Storage (MCS), designed to close the performance gap between storage and compute. A key use case is GPU Initiated Direct Storage (GIDS), where data moves directly between storage and the GPU, dramatically cutting system latency.
The company is now developing seventh-generation Z-NAND, the key medium behind MCS. Its third-generation product, tailored for GIDS, is expected to deliver throughput well above current industry benchmarks while maintaining ultra-high performance and peak energy efficiency. "Memory Class Storage will reset expectations for AI inference speed and underpin real-time intelligent applications," Kevin Yoon emphasized.
In closing, Kevin Yoon stated that through full-stack innovation in memory and storage, Samsung not only aims to solve AI-era data management challenges but also unlock the full potential of AI technologies. Moving forward, Samsung will continue to push technical boundaries and work with global partners to build an open storage ecosystem and co-create an intelligent future empowered by AI.
Honored for Excellence: Outstanding Storage Technology Leadership Award
The GMIF2025 Innovation Summit, in addition to in-depth discussions of technological evolution and ecosystem collaboration in the AI era, also announced an award list that highlighted outstanding enterprises, innovative technologies, and exemplary solutions that have emerged in the storage field over the past year. As an industry leader, Samsung Semiconductor has continued to make major strides in both DRAM and NAND technologies and was honored with the "Outstanding Storage Technology Leadership Award."
Samsung Semiconductor honored with Outstanding Storage Technology Leadership Award. Credit:GMIF
The judging committee noted that as AI fuels surging demand for high-bandwidth, energy-efficient storage, Samsung has moved fastest in bringing advanced technologies like HBM, DDR5, LPDDR5X and high-stack 3D NAND into large-scale production and real-world use. These innovations have helped accelerate large-model training, data-center upgrades and next-generation smart devices, giving global customers a much stronger technology backbone.
The award also reflects Samsung Semiconductor's long-term investment in R&D and its role in continuously pushing the limits of storage performance and enabling digital transformation worldwide. The company reaffirmed that it will keep driving memory innovation, deliver cutting-edge solutions across the ecosystem and help advance AI adoption globally.
GMIF2025 brought together top industry leaders and the keynote by Kevin Yoon offered a clear view of where storage technology is heading in the AI era. From agentic AI and breakthroughs in GDDR7 and CXL to progress in high-density storage and thermal design, Samsung is clearly at the forefront. Its introduction of Memory Class Storage adds an entirely new approach to AI inference. Moving forward, Samsung will continue working with partners around the world to unlock the next wave of storage possibilities and power an AI-driven intelligent future.
Article edited by Joseph Tsai
